One-step ahead modelling of building thermal dynamics: A comparison of backward selection and LASSO approaches

The computation time and the infinite possibilities in model structures of data driven models often hinder the efficient development of accurate models. This paper presents a systematic approach for selecting the appropriate model when forecasting the temperature dynamics in non-residential buildings, using different streams of data. The main objective of the work is to evaluate the presented approach by comparing the results to those obtained by a typical backward elimination method. The workflow delivers the selection of the appropriate features in order to represent the system accurately in a parsimonious model, by setting the initial model structure search space and then estimating the parameters, using the least absolute shrinkage and selection operator (LASSO) procedure. The analysis is performed on a case study educational building complex at Loughborough University in the Midlands, UK. The input data comprise of multiple features including internal room air temperatures, external air temperatures, and HVAC related data such as valve positions and fan speeds, measured sub-hourly over a winter period between 2018 and 2019. The results confirm that the specified automated workflow enables accurate estimates of indoor air temperature using considerably less computational effort than the backward selection approach. However, the final form of the models identified could lead to poor control performance.